Latest / The Information Bottleneck / Kaggle Grandmasters, Agent Skills, and Why Everyone Is Overfitting with Jean-Francois Puget (NVIDIA)

Transcript

Kaggle Grandmasters, Agent Skills, and Why Everyone Is Overfitting with Jean-Francois Puget (NVIDIA) · The Information Bottleneck

• Ravid Shwartz-Ziv: Hi everyone and welcome back to the Information Bottleneck. And today we are going to talk with Jean Francois. He is a director and distinguished engineer at Nvidia. Hi Jeanne. Thank you for joining us. Jean Francois Puget: Hello. Thank you, Ravid, for inviting me. Ravid Shwartz-Ziv: And hey Alan, thank you for joining us too. Allen Roush: Hi Raveed and hi JFP and I'm really excited to chat with you about Kaggle because sometimes I I forget that that place exists and quietly makes some of the most skilled AI engineers stand out. Ravid Shwartz-Ziv: So today we are going to talk a lot about Kaggle competitions and also about agents and how to use agents and how NVIDIA thinking about how to use agents. But maybe we can start with the recent announcements from NVIDIA this morning. can you say a few words about like how NVIDIA and how you think personally about agents? Jean Francois Puget: Yeah, OK, so this morning, Jensen, our CEO, made a number of announcements at Computex in Taiwan. So I won't cover everything. He made some announcements on hardware, like chips for laptops. But I'm not on the hardware side. I'm on software at Nvidia. So closer to me, there are two main announcements. One is a new larger version of our NemoTron model, NemoTron Ultra, which is comparable to top open source from China mostly. But the main difference is it runs much faster. Tokens per second output is way higher. because this model has a hybrid architecture where some parts of a classical transformer were replaced by state-space models like Mamba. So that's one. And the other, it's optimized for 4-bit quantization, which is something NVIDIA is pushing both in software and hardware. So we'll see how people consume this model, but... you get similar performance like Quen, Adipseq, kind of models, Kimi, but much faster. So we'll see. Then a little caveat, you have numbers and benchmarks and you have reality. So we are looking for feedback from reality, but on benchmark it looks very competitive. Then closer to what I'm doing, NVIDIA also announced an agent skill repo with skills for NVIDIA software. So agent skills, let me say agents coding agents like Codex, Cloud Code, Cursor, and open source like open code. They are great. You can do a lot. I'm using them all the time. But they may not know very well some NVIDIA libraries. They are trained on what you have on GitHub, for instance. And maybe they are not a lot of examples, especially for new libraries. So you need to help agents. And you help them with documentation, tell all to them. But there is a specific type of documentation called agent skills, where you document how to perform a workflow, how to... accomplish a task. So a scale is just a prompt in markdown format that you would give to an agent if the agent needs to perform a given task. So in a scale you describe the task name, what the task is, and then the prompt to follow. And this has emerged as a standard, started with cloud code. And now it's a standard supported by major agents and not only coding, but also PUNCLAW or Hermes agents. And at Nvidia, we wondered how to expose our libraries using agent skills so that generic coding agents or generic agents can use our libraries effectively. So we just shipped on GitHub Nvidia skills. number of skills and our angle so creating skills is easy. Creating useful skills is a bit more tricky. And I've tried a lot of skills. even the first at Nvidia to speak about skills. And I've tried a number of them. And surprisingly, a lot of them are not helping. Frontier models in these coding agents, the codecs, but now there's, they work well. So improving, honestly they are perfect, not at all, but they work well. So improving them, you must show you improve. So I've driven an activity to define an evaluation standard for NVIDIA and a way to test. So basically you define a set of tasks, you want your agent to perform. and then you compare how the agent performed the task without the skill. So it's quite simple, but just doing this consistently, reporting on the results, automating this test, we now have at NVIDIA a pipeline where people can just submit skills, have them tried with codecs, codec, cursor and others. and have a report out of it and maybe recommendations on how to improve the skills. So we've automated evaluation. And the one that passed the bar, the first ones were just announced and there will be more. And this has triggered interest from third party like OpenClaw and Hermes and the big frontier labs so that they will redistribute our skills as well. And that's something we have done recently. Again, the goal is to help agents not trained on NVIDIA software to use NVIDIA software. Ravid Shwartz-Ziv: And how much like at the end how many skills you think there will be that related directly to to NVIDIA? Do you think like it's hundreds of skills? Jean Francois Puget: Yeah, more than a hundred, yes. Allen Roush: Do you Jean Francois Puget: And it's increasing every day, so I don't have the count now, but I've seen a counter 121. So I don't know how many are published yet, but the pipeline is hundreds. Allen Roush: Do you have any opinions on the value of so for example, there's different implementations of skills, like clawed codes skills are are prompt, as you've mentioned. But Codex, I've seen at least one skill that supported executing like deterministic specified code. And I think MCP servers are often what people imagine skills to be. You know In a sense, in in in that they connect people up to data that enables their models to do better. So do you think that these like data connectors can have more value than the skills or or what's your opinion on that? Jean Francois Puget: No, no, no. First, there is no difference between skills used in codecs or cloud code. You can use one with the other. It's the same stand out. But the skill is a skill.md file, which is markdown. So a name, a description. You can have a bit more in the header in the front matter like license and whatever. And then you have instructions, so a prompt in markdown. And the skill can refer to code, to MCP tools, to whatever. So, or CLI that the agent should call. The point is, if you have a MCP tool and you want it to be used in an agent, you need to give a description. And the description is loaded whether the agent needs this tool or not. So the problem with MCP tools is it eats your context very quickly. Agents, kids have something called progressive disclosure. So in theory, it doesn't work perfectly, but in theory, When you have a set of skills, the agent will only read the front matter, the name and the description, which is a small part. And if the agent decides it needs the skill, it will read the rest. So you would put the long, then you can create a skill for your, for each MCP tool. You just have the, the front matter describing quickly the relevance, and then you have the detailed instructions on what the MCP tool does and examples, et cetera. And this way, you don't eat your context, is loaded on demand. And that was the first selling argument of skills. People were reporting that moving from MCP tools to skills, they were reducing their context consumption by a hundred times. So that's huge. now if you want an agent to use a CLI or a data connector, you can encapsulate with an MCP tool, but you can also just document how to call the CLI in a skill. you, the agent will just say, okay, if I need to have to use a CLI for, I don't know, Kaggle for instance, I can use this skill, but I know there is a skill about Kaggle CLI. And then whenever the agent needs to actually connect to Kaggle, it will read the CLI description in the skill and it's loaded only on demand. that's, so you, one way to think of skills is to have a modular prompt. You load the part of the prompt you need when you need it. So that's the array in practice and that's why we need to test. So among the default metric we use is. Is the skill loaded when it should be and is the skill not loaded when it should not? Both are issues too, so we measure both. So we have a set to test the skill. As I said, we have a number of tasks. must include the test set must include tasks that do not need the skill so that we check that the skill is not loaded when it should not be. yes. Ravid Shwartz-Ziv: And and do you think like in the future like it will be like every like now that I every like company or like I don't know like some setting that I want to adopt, like I will have like just new skills or do you think like the solution will be to I don't know, fine tuning, go like doing something more permanent? Jean Francois Puget: That, okay. That's maybe, so let me give an answer. Maybe that's not exactly your question, but now if you have an agentic system and you want it to learn or evolve, you can update the model, save your reinforcement learning. You can evolve the prompts. So there are some methods like JEPA or DSPY. Not used a lot, but maybe they are underused. Change the agent itself, or you can just add more skills to your agent. And that's something we, the latter is just, you add more skills. So you just add files and the agent just use the additional info automatically. So we want a benchmark called Dabstep, which is a science benchmark. And we want it by, so we had a data science agent. We made it solve each training data, training. task in the benchmark and for each of the tasks the agent looked at how it solved the task then created a skill out of it. So it accumulates agent skills. So for each task you have a skill describing the workflow used to solve that task. So it's just taking the agent trajectory and that could have... been too long and for the trajectories that succeed. So maybe there were more than one trial, but for the success one, then this still the trajectory into a skill. So a problem that tell the agent, do this, this, this, and this, and you solve the task. And then, so if you have a hundred tasks, you have a hundred skills, which is too many. And then a second phase, we had the agent. look at the skills and generate a smaller number of more generic skills. And then at the end, we use these more generic skills with the agent to work on the test data. submitted this for the test and we won the benchmark that way. So it's a new way of learning. The agent accumulates procedural knowledge in markdown files. So we don't change the agent architecture. don't change, we don't fine tune the models. We just. add more context for the prompts in the agent. And this is quite effective. I believe it's a new form of learning. We're not the only one having this idea. Independently, NUS Research, for instance, who develop Hermes Agents, they build their agents on this as well. Their agents continuously create skills as it goes and They were lacking the compression part that we had. So they've added it a month or so ago. As you relate, you need some time to distill them into a smaller set of more generic ones. Yes. Ravid Shwartz-Ziv: But you don't do. But you you don't afraid from kind of like overfitting because like at the end if like w when you see more like when you create more skills and like even if you compress them, right, like still you can say, okay, like the the model at the end will say, Okay, so like for these very specific types of problems I will decide this and like for these very specific types of problems I decide that. Jean Francois Puget: Maybe it's so first I'm glad you're asking about overfitting because that's I love it when people discuss overfitting. It's a great topic. That may be true. the compression is really the goal is to make skills as generic as possible, but could still overfit. But now if you think there are two flavors of long running agents, one is you have one task like find a new molecule or solve a math problem. And you will want to, it's fine if it runs two weeks on a single task. And then there are long running agents, especially personal agents that runs repeatedly similar task. Say you have an agent that, I don't know, that just your email and Slack and social media and blah, blah. If it generates and then you tell it when it did a good job or not, and then it generates skills to overfit to what you believe is a good way. Overfitting is fine. It overfits to you and that's okay. So overfitting may be fine if you never use it outside the training data distribution, right? So it can be good in that respect. But yes, you're right. That's why the... Your question about overfitting, let me return it. I believe most LLM benchmarks do not check for overfitting. Overfitting is rewarded. Say you take a benchmark, test data is public. So you can say, let's take MML UI. Ravid Shwartz-Ziv: What do you mean? Jean Francois Puget: I'm not saying it's a bad benchmark, just an example. Say you fine tune a model on mass task, you run it on MMU, you have a score. If you use this score to guide you on your checkpoint selection, you overfit. I don't think so. I've even seen people publishing papers to improve score across a set of benchmarks automatically. That's overfitting by definition. Ravid Shwartz-Ziv: Do you think it's a problem? Jean Francois Puget: Yeah, it's a problem and that's why people continuously report that in real usage, performance is not as good. It's always so. Ravid Shwartz-Ziv: But like I I will people will claim, right, that like currently with like in the LLM area, right, that even though like the like most of the of the benchmarks have like public testing set, still if you improve on these things like everything will go up. Jean Francois Puget: I agree. So the benchmarks are good to run the models, but they are not good to estimate the actual performance and actual problems not seen before. What I'm saying is there is a drop between the benchmark values with public test data and actual use performance. Ravid Shwartz-Ziv: Okay. Allen Roush: Well Jean Francois Puget: What you say is if a model is better on benchmark than another model, it would be better in practice. This is probably true, but it does not change the fact that both overfit to the public test. And a great example, so I'm sidetracking, but I want to say before I forget. Last year we worked on sweet bench quite a lot. Sweet bench verified. And our team at Nvidia won the benchmark at that point. We were on top. We stopped working on it because we tried 03 at the time. And we found that 03 was, our agent, so Swaybench, it's about solving GitHub issues. you're given a repo, a tag. So you check out a repo, you have an open issue on it, and you need to produce the edits in the repo that fix the issue. The problem is that these edits are in a commit in the public repo as well, so LLMs are trained on it. So we found that, so our agent was first trying to identify which files in the repo to edit and then do the edit straight forward. We found that O3 was picking the right files to edit without looking at the file content. Okay, so what's the meaning of the benchmark after this? So my position is the only reliable benchmarks are benchmarks where the test data is not public. So it has to be managed by something. And if you want to evaluate, you train on the training data and when you're ready, you send your... Checkpoint, URLM or whatever is tested to whatever runs a benchmark. But this is really a minority. Ravid Shwartz-Ziv: But but even here like you have the problem that like even if you don't have direct access to the test data, you still get like some signal, right? So you can optimize over this signal. Jean Francois Puget: You have one shot, you don't loop, you can't loop. I agree if you can loop and use the score, you defeat the purpose. But the difference is, so you have to choose what you submit before you know the score and not after. That's the difference. And that's how Kaggle competitions are built. In Kaggle, know the score after you've selected the two submissions you make for the final scoring. And that's why it's valuable. It's not the only place, but I think it's a good model for benchmarks. Allen Roush: Yeah. So So so I just had some some questions about about Kaggle. and and also in relation to Evals. First of all, Kaggle is something that long predates the you know post-2022 Chat GPT generative a era, and was r can continues to be a really big deal for kind of figuring out who actually is good at at least, you know, training classifiers and other optimizable kind of verifiable rewards. know, in machine learning. So, I guess my my question is how has that shift to language models really shaped Kaggle in the last three or four years? And then also just a c a quick side question around you know, we you we you two have been talking about how you val you know validate models. I mean back in the era before chat GPT, back when people used Stack Overflow, there was a slightly lesser well known Cross validated being the name of it, machine learning variant of stack overflow. And cross validation, that technique, was something that was really drilled into my head in machine learning classes. Jean Francois Puget: Yeah, so let me start with this next story. Ravid Shwartz-Ziv: I will before you answer, I would just like maybe we didn't say it like you're a grandmaster at at Kaggle like many times and like you had like a long history on Kaggle and like you're running a team that actually like has like again expertise in Kaggle competitions, so yeah, now you can answer. Allen Roush: Yeah. Jean Francois Puget: Yeah, I'm very good at it. Allen Roush: That's why we're asking this stuff. Jean Francois Puget: Before joining NVIDIA, I was asked, well, I did a PhD in machine learning a long time ago, then moved to something else, mathematical optimization. And we go to work by IBM and someone noticed I have a PhD in machine learning, which is irrelevant now. It was very different at the time. But they help us on machine learning and I say, where can I find? current machine learning practice. And I found Kaggle and I say, guys, I have a PhD. You will see what you will see. I entered and I didn't do well in the first competition. So they're okay. need to learn. And I became good at it. As you say, there is a... Yeah, a ranking of the greatest of all time on Kaggle and I've been third on it. So I've been active and that's how I entered NVIDIA actually, thanks to what I do on Kaggle. And yes, to your point, Alain, one thing I did at the time as well was to take Andrew Ng training on machine learning and He made it clear that overfitting was the evil trap for machine learning. And cross validation is the main weapon for it. So I've integrated this. I did bad in my first Kaggle competition because I overfit. And that's the only time. I have a reputation to not overfit on Kaggle. I'm very conservative. This was, well, that was the difference between a good data scientist and a bad one. It's not the good ones don't overfit. And to your question with ShadGPT and now GenAI and Agents, all this knowledge is disappearing. People just let fusion do their job. And I'm seeing it on, so on Kaggle, Back to your question, how is Kaggle impacted by JAN-AI? Kaggle compute is limited. You cannot use frontier models. So cost would be too... So they cut internet access, one for the cost, so going to not leak the test data. So you cannot use frontier models when you submit on Kaggle. You can only use small open source, say four billion or eight billion maybe. So that's, you don't use elaborate GenAI inside the submission, but you use it outside to prepare for your submission. And there are people use agents, they use codec, they use cloud code, they use whatever, they use frontier model to run agents continuously or to generate synthetic data so that, so there is a lot of opportunity. for pre-training, whatever you submit to Kaggle. So the shift is in automating the work of the data scientists, basically. So there are a number of competitions where all top teams claim to use AI agents. I predict there will be a bloodbath because they overfit, but who knows. because these agents, they have no notion of overfitting or they are told to improve the score, they improve the score period. So we'll see what happens. When is ending in two days? I'm looking for it. Ravid Shwartz-Ziv: Which one? Jean Francois Puget: Berkeley. So it's a competition I did every other year. It's like the fifth edition and I won gold medals in every of them. And I know the main issue is not to overfit. So we had to be very conservative. I doubt all the vibe coders or vibe data scientists understand how how much overfitting is easier now with agents. Ravid Shwartz-Ziv: So so what do you think like what do you think actually like make like a good data scientist, you know, like for Kaggle competition. Now what do you actually what are the types of like problems and solutions that you are trying to to solve and how? Jean Francois Puget: So I believe what makes a good data scientist is two things. The first and foremost is to be able to evaluate the performance of a model accurately, say via cross-validation. Because you need to select your model or pipeline before you know the score. So you need to be able to evaluate your model correctly so that you pick the right... So during a competition, you train a number of variants and you have to pick two that are scored on a hidden test set by Kegel. If you don't pick the two right, your two best, well, you don't do as well. And usually competitions after completion, people... The forum is full of people who say, look, had this variant that would have won the competition if I had picked it. OK, but you didn't pick it. So you missed the point. So evaluating properly, that's a key thing. The second one, but this is changing, to do well, you need to experiment in a scientific manner. You have an idea, you must think about how to evaluate if it works or not. And then you perform the experience with and without the idea. And depending on the result, you conclude that the idea is useful or not. So it's like any experimental science. So you have to do it rigorously. and efficiently so that you perform. The more experiments you do, the better in the end. But you need to evaluate properly. And I see people, for instance, they modify five things at a time. And then if the result is not good, what do you conclude? Maybe you had two good ideas out of the five and three that were bad. So just change one thing. An experiment gives you one bit of information. It's better or not. So you should only test one thing at a time in an experiment. You cannot test several changes. And this may sound obvious, but I still see it all the time when I team with people. They create a new variant. They don't even run cross-relation. They submit. it's not as good. Okay, but you change two things. What do you conclude? That both are bad? No, you can't know. So if I summarize, evaluate properly, and then perform experiments as fast as possible. Where agents are changing, the game is you can use them to automate experimentation a lot. So I have people on my team, they run Codex or Cloud Code continuously and it runs experiments in parallel at large scale. And this can be really large scale like running experiment on slurm clusters for you and continuously. it's a... It's leveling a bit the field. It changes a lot. And as a result, my team is the top consumer of tokens inside NVIDIA. Ravid Shwartz-Ziv: Do you have leaderboards? Jean Francois Puget: Well, if you use Codex Enterprise plan, you can have a leaderboard of users for your organization. And the top three are from my team. Allen Roush: Wow. Actually related to that question, there's also the Auto ML revolution that I think was happening right before Chat GPT and LLMs got really good. And there were pr pr projects that I'm sure you're familiar with, such as Auto Gluon, which I think comes out of AWS, Auto Teapot, and etc. Many of them feel like they're on life support and even whole startups like Jean Francois Puget: Yeah. Allen Roush: Ding. the like a whole Auto ML startups I feel like kind of died during that time. What's your opinion on that? Jean Francois Puget: and because... Yeah. So automating experiments with the coding agents is killing this. We get better results. Allen Roush: H two that's what I was thinking, sorry. Jean Francois Puget: Yeah, that's good. Ravid Shwartz-Ziv: So but but okay, but but do you think like the the future is like coding agents that are doing this like hyper parameter sweep and like running exper experiments on the clusters and you just like set the right metric with the right validation? what about like creativity of ideas, for example? Do you think agents are there or Jean Francois Puget: Sure. First, I still have my doubts about overfitting. So we'll know soon because there are number three, four, K-gallon competitions where people just rely on coding agents. So we'll see. If, as I believe, overfitting is an issue, then that's an area where the human controlling the agent should work. That's so training a classifier or regression model or pipeline is well understood. So I wouldn't be surprised if this gets automated quickly. But then where you need creativity is for more complex domains. I don't know, drug design or general agents navigating an unknown environment or. So I believe the frontier, I believe LLM and agents will digest repetitive tasks. It's automation. But they are not at the level of human creativity. How can I say it? So I'm using codecs lately, most of the time. And my view is people say, you should treat this as a junior intern. No, because it can do things that a junior intern can't do. It can do a lot very well, but it can be wrong by large proportion as well. So it's a mix of a genius and a dumb person. It's extreme. When it works, it's awesome. And when it, and sometimes it makes mistake, no, no human would do. It lacks common sense. It the limit. It does not have a global model of what the code does. So it has a local view, very accurate, but I keep arguing with my coding agent. I ask, okay, update and do this and this and this. Yeah, but there is this assumption, this assumption, so we must test for it. I say no, because in this other part, we live in a state where the assumption is true. yes, you're right. So it cannot, it lacks a system-wide understanding. And when I see people who let coding agent run for hours or overnight. I think that what they code is simple. My experience is when you have something complex designing new algorithms or... with system-wide dependencies, it's complex. It's beyond what they do. For instance, an experiment colleagues did. They created automatically in two months an equivalent of PyTorch. So the fact that this could be done is impressive. There was an NVIDIA blog recently, like two months ago on it. So a fully autonomous multi-agent system created an equivalent of PyTorch, which was functionally correct. That's impressive. but it's 10 times slower than by touch. And it's because there is no system wide optimization. Locally the code is great, but... Ravid Shwartz-Ziv: So but do you think this is something that will be solvable in the future or do you think they could keep Jean Francois Puget: Thanks Everybody who said, no, they would never do this or that proved wrong so far. I'd be cautious. What I'm saying is a current limit I see to coding agents is they lack a global context. So that's where humans still have a role. So I still believe humans are in the driving seat. But it's like mechanization, know, in the past, they were using their muscles to work with the ground and the soil and do stuff. And now they have machines and they driving the machines. It's the same for agents, it's automation. But the machines, they work both better and not as good as people say. When they work well, the agents do really, it's impressive, the code they can write quickly, etc. But you have to check because sometimes the code is just wrong and I don't understand how people can trust it blindly. They shouldn't. Allen Roush: So, the the the the work that that you're doing right now with these code agents, you know, and and talking about many of these different harnesses, I'm aware of like Nemo tr the the Nemo Claw, and my understanding was that Nvidia p some of what Nemo Claw's trying to do on top of OpenClaw is improve the sandboxing for enterprises. Yeah, I just wan I wanna hear your opinion on sandboxing in in general for for models. Jean Francois Puget: Yeah, so that's another issue indeed. So we see our stories. There was one a month ago where a company... I think it's right weight. They lost all their code or all their, I don't know, because the agent erased some database. So my take is there should be always a human in the loop. I don't believe full automation is okay, except if you make it solve a task and it iterates till the task. But you look at the end result. So as long as it iterates, it's fine. If it goes wrong, too bad for you, but that's it. But consuming the output of an agent without... looking is wrong in my view. But people do it anyway. So you say, OK, should we sandbox? So Nvidia answer to this is called Open Shell. And we announced OpenShelf for OpenClaw, believe, this morning, and maybe Hermes Agent, I believe both. So it's a way to sandbox. So sandboxing is controlling. Mostly about controlling. actions and outbound communication but it's not perfect and then Nemoclo is an installer that installs OpenShell with your favorite agent but the sandbox is called Hermes so my view is that we're just starting a long journey about securing agents What I believe is it's extremely... So coding agents are good at automating software engineering, but they are good at automating hacking as well. And recently in France, a hacker was like a 15 year old was arrested and the guy in six months hacked a number of sites and... became famous in hacking community and not just in France. And when he was arrested, he said he did not write a single line of code. We use coding engines to do all the help. Allen Roush: And that was way before Mythos. Ravid Shwartz-Ziv: So What do you think about Mytos? do you have any opinion about it? About like the you know to release it to the public versus not? Jean Francois Puget: Wow. I think it's a capable model, but I dislike anthropic communication, not just on this, but on the rest. In the past, no, it's a bad one. Years ago, Panayai was a bad PR. They were announcing semi-conscious models and AGI next quarter, blah, blah. They have stopped. Ravid Shwartz-Ziv: The best PR machine. Jean Francois Puget: And now on Tropic is fear mongering all the time. AI will eliminate 10 % of jobs in the next six months or 90%. I don't know. It depends on the day. Mitos is too dangerous to be shared. We must forbid Chinese models because they will be used to hack, blah, blah. So they try to kill open source, try... If Mythos is so great they shoot open source and that everybody, if they have vulnerabilities, themselves... You know, if all they want is to benefit humanity, that's what they would do. But of course, that's not what they want. They want to create fear so that people que to use mythos presented as the only way to avoid say, you don't use mythos, someone else using mythos will kill you. We know this kind of argument. So the good news is... Ravid Shwartz-Ziv: Yeah. Jean Francois Puget: OpenAI is reporting similar performance, at least on public benchmark, and it will take a few months before open source achieves similar performance. So they have a small time window. They try to monetize by telling people how great Mitos is and how dangerous it is if you don't use it. So I believe it's a good model for what they... So technically I think it's a big progress. Ravid Shwartz-Ziv: W wh what do you think like are the missing parts for open source models? You know, like why Jean Francois Puget: don't think there is anything bad there. Lots of reports show that they trail frontier moor, a close source, by six months or four months, or it depends on the report. So they don't miss a lot. They are coming. So what's demo data achievers at a point you can embed that within a year, you can do the same with open source. Ravid Shwartz-Ziv: But But do you think it's like what, like that NVIDIA do doesn't put like enough like compute, enough people, like d don't like I don't know with data? Jean Francois Puget: Why you say why is not a frontier model? So first, in some areas, we are the frontier, like speech to text, Nvidia is really among the best. Now for, same for image and video generation. So we demonstrated what we call world model to generate. actual videos that can be used to train autonomous vehicles. So that's a big activity at Nvidia and probably Tesla something great there as well but well Frontier and now it's general tech generation where we are not Frontier or not OpenArea or Anthropic or even Google. But it was decided to, last year, to accelerate on open source and that's the Nymotron you'll try this morning that shows it. So it's competitive with the best Chinese open source and Nvidia will continue pushing. So expect better models from Nvidia, open source models. Maybe they will always be a bit behind frontier models, but not so far. So what do people need? Me, sorry. I believe something and it happened in previous life when I was in mathematical optimization. If you have the best team in the world on a topic, the best people not on the team wants to join you. No, that's, that's, I behave, so it, being top in a field attracts the top people. And I think that works for Anthropic and OpenAI these days. It's a dream, it's a dream of most, not all probably, but most people working on LLMs would love to join one of these because you know, you're part of the best. So I think NVIDIA focus on training frontier model was not known much and well not yet at the top. But we are hiring great people all the time now. I can tell you not just for this but for agents and reinforcement learning and another area where we are. The front is for robots, so real world physical models for robots. We have Groot, I'm not a specialist, but it looks like one of the strongest models available. At the same time, the goal of NVIDIA is not to replace Anthropic or OpenAI. are, especially OpenAI, one of the best customers. So it's more for people who can't afford the top to give a good fallback to the open source. And there is also another reason why people can't use frontier model necessarily. It's when you apply it to confidential data. Ravid Shwartz-Ziv: So like w w how do you see like the the future? Do you think like we will still have like I don't know, two or three frontier models that are closed source and different open source that like if now you have some confidential data or like you want models to run locally or to fine tune on your own data or like something different, I don't know, very like specific models for specific domains and Or like maybe agents that's using different models at the same time. Jean Francois Puget: I think one evolution I'm seeing is what people call world model. So far we have models that are good at interacting with software. But if you ever want to have a robot that, you didn't know, but it's a classic definition of it. You know, I will personally be convinced that we have an artificial intelligence. Literally is when I can ask a robot to make me coffee in a place it does not know. you know, looking for coffee machine for beans, if not go to a shop, buy coffee or, you know, all sorts of things that are not rocket science at all, but that you would do naturally as a human. You know, you're you asked to do a task and you will find a way to know mostly to solve it. There is nothing in current research that. is ensuring progress on that front. So I believe we will have systems that can perform specific tasks as well or better than humans in many places. And this is not new. mean, we know AI, not the same flavor as now, but AI is better than people at chess, at Go, at games, at... As I said before, I worked on mass optimization. It can solve planning problems better than anyone, et cetera. So this will happen. But having a autonomous general purpose system. You know, this is not a focus, but I see more and more people starting labs to work on something a bit like this. So more autonomous and more general purpose. And Leucan, for instance, is one, but not only. are US labs as well being created. Moving beyond. Ravid Shwartz-Ziv: It'll take time. Jean Francois Puget: what LLMs and agents using LLM can do. I have no clue how far they can go and how fast they will go, but it's starting. Regarding LLM agents, I believe this is going to automate a lot of white collar task, which does not mean replace people. It means that people will... Ravid Shwartz-Ziv: But Jean Francois Puget: will have AI work for them and they will focus on something else. Ravid Shwartz-Ziv: And do you think like at the end the solution for like I don't know, end to end or like real world and problems will be different than Kaggle competitions? Do you think like fundamentally we need different solutions for these problems or we can transfer and like we can learn on one very like specific setting and then to transfer these solutions to to another setting? Jean Francois Puget: So in Kaggle you have two current competitions that... are about agents acting in a complex dynamic environment. So one is orbit walls. The other is ArcGIS-3. So these are steps you want to develop autonomous agents. that can make a decision real time in a dynamic environment. But if you look at, so this is going in the direction beyond what LLM can do. But if you look, these environments are extremely simplistic compared to the real world. It's really, as I said, it's really the starting of a long journey. I don't know if Kaggle will be the relevant place. I believe to move forward, you need more elaborate simulations than the ones in these competitions. And for instance, if you look at DeepMind, they started in making agents for games, like... So, AlphaStar, I don't remember, it was Dota maybe, so a complex multiplayer game. The environment was more complex than in these competitions. Maybe the way to get to the kind of autonomous AI I define as AGI, you need to work in video games, or I don't know, which are more more realistic to experiment. Because in video games you don't need to the physical robot. So it's easier. even if you have AI that's autonomous and learning this environment, then you will have to put them in a robot. And that's another level of complexity. So the journey towards autonomous robots is long. the terminator is not going to happen anytime soon. Allen Roush: I'm fascinated that you brought up Dota. I I'm I'm the gamer. Well, I I think Ravide plays a lot of games too, like retro gaming, but I'm definitely the one who's currently the the gaming addict out of the two of us. And I'm I'm always fascinated, do you do you think that the relative weakness that LLMs or AI agents have in playing certain games, things like Dota and also like NetHack, which is also well studied and famously difficult of a game, do you think it just goes back to like the context window and agent planning and agent communication, or do you think that there's kind of other factors or even like maybe an architectural problem where we need like some rethinking outside of like the L pattern? Jean Francois Puget: Yeah, first, I believe it's not seen as a major challenge. So you don't have people spending billions on it, on solving it. People are spending billions in adapting what seems to work for math coding to other domains. Remains to be seen if they will succeed, but... My take is we need something quite different to do well in dynamic environment with multiple real-time decision making. But I may be wrong. I don't know if LLMs can produce tokens in the millions per second in few years. Who knows? I have no idea. My gut feeling is we need something different. But the story has shown that so far transformers are eating the world. And it's because they are done. So we'll see. Ravid Shwartz-Ziv: and so we're almost out of time but I have like one more question. Like d w what do you think about the this claim that like at the end AI is kind of a bubble and like the the the only company that actually makes money out of it is Nvidia. Jean Francois Puget: What I believe is some people make too bold claims. I said, OpenAI was, but they stopped. Now it's entropic. So this can create expectations and then people are disappointed. So there is a risk. I'm old enough to have seen an AI winter and it happened for the same reason. People are just making too bold. And there was an AI winter even before I was born. In 1960, Herbert Simon and Alan Newell made predictions in 1960 that in 10 years AI would be the world chess champion. So it happened, but it happened like 30 years later, not 10. And Antropik showed Arihama Day prediction that AI will replace 90 % of wet-cooled clubs. This is really not helping. So that's for the bubble part. Then the bubble. No, I really. I think what we will see is as the token cost decreases and it's decreasing every few months, it will drive adoption more. If you look at people who tried, who started using Cloud Code and Codex and others, they're hooked. Going back is difficult. Same for Chagypti, people who started using Chagypti or alternatives now. Going back to not using them, it's difficult. So I don't think it's a bubble. Then there were arguments about circular economy, blah, blah. If you look at the level of investment and the money in the other way, it's not circular. And not just for Nvidia. Google invests in Entropiq or Amazon and they sell chips. Nobody says it's circular. So I don't think there is a bubble. I don't understand. As Jensen said, the way Nvidia stock price evolves and is due to the fear of a bubble, probably, he doesn't understand. Nobody understands that. must say, what are people afraid of? It's not clear. So we'll see. I don't believe it's a bubble. I believe some people make too bold claims. So if people Trust these claims, then yes, it's a bubble. But, you know... This is not just in AI. mean, there is a company claiming they will establish humans and Mars not too far in the future. You see a kind of claims as well that we may not see anytime. So people make bold claims. I don't know why. don't know why. Probably for funding. Ravid Shwartz-Ziv: Okay. w we we didn't talk about ARK prize, maybe briefly what like you had like you you competed in the this competition. Maybe tell us in a few words about it? Jean Francois Puget: Okay, so... Yeah, so... Okay, yeah. So RKGI is a set of challenges created by Francois Chollet, who's the inventor of Keras. He was at Google Demind for a while, now he has a company. So RKGI 1 and 2 were about solving puzzles. So you were given, you have tasks and for tasks you were given some example of puzzle. So puzzle is... you have an input, a little grid, and then you have an output and there is a transformation like you color all the interior of points in yellow. So that's a transformation. So you had to build a system that learns from very few examples how to move from input to output. So it's about few-shot learning, really. in a little visual environment. And last year when the competition started, Frontier-LLMs were solving like 1 or 2 % of the tasks. and one or two percent, yes. And we won with a four billion model on Taggle compute. So meaning at a cost of two, of 20 cents per task. We solved 27 % of the task. This was the best score available. Since then, the frontier models are... Ravid Shwartz-Ziv: Book two. Jean Francois Puget: have made progress, but so far nobody beats us on a Pareto frontier. So nobody does better than us at the same cost. So it was a huge, huge achievement because what was incredible is before the last day we were third and we jumped to first place and increased our score by almost two X in the last day, at the last hour in a seven month competition. Ravid Shwartz-Ziv: How? Jean Francois Puget: Yeah. And how did we do it? So to solve this with a 4 billion model, it's impossible. So we actually distilled frontier models and frontier open source like GPT-OSS and road 20 at a time. So we looked at the puzzles. We use LLMs to generate additional puzzles. So we used code generation and the capacities of frontier model reasoning to create a set of one million puzzles. And then we trained our small model on it. So we distilled frontier models on this benchmark. And there is a new iteration of the competition and we are leading it. We are nearly at 50 % now. still at 20 cents, which is huge compared to what frontier models. To get 50 % with frontier model, you have to spend dollars per task instead of what we do now. And then there is an ArcadeGI 3 where it's not static. It's like small video games you have to solve. So it's not Dota, but it's moving into, you have opponents. This one is very challenging and the best score on Kaggle is like 1 % now. So it's really challenging. Ravid Shwartz-Ziv: So and w what do you think about like the the arc GI three? Do you think this is something that it's solvable? Do you think it's something like we Jean Francois Puget: I it's very hard. We are entering it, so we'll see in a months. It's a several-month competition. If you are not limited by Kaggle and you can use frontier models, people publish multi-agent system that solve up to 40 % of the tasks, so it's doable, but at a very large cost. the question is, can we have a small system do something? It's open. These are well aligned with what I say, exploring the limits of what you can do with LLMs. these are good. This is way past regression and classification. It's really... So I hope for Kaggle they will have more competition like this. Time will tell, but these are relevant. Ravid Shwartz-Ziv: Okay. I think we're out of time. do you have anything else that you want to add? To advertise? Jean Francois Puget: No, no, no, it's for people who watch this. None of this was prepared. I didn't know about the question. So I hope it's interesting and entertaining. and I would if people want to see really what SOTA is in practice, just enter Kaggle competitions. To me, it's a legal drug. Kaggle is a legal drug. So it's legal to do it, but it can be time consuming. But it's worth it. It's worth it. Ravid Shwartz-Ziv: I agree. okay. Thank you so much for joining us. It was a real pleasure. And thank you for all the audience. Thank you so much. Jean Francois Puget: Take me. Thank you. Allen Roush: Thank you everyone.